Information Technology Reference
In-Depth Information
User affinity graph W
U
. Generally speaking, the activity of joining in interesting
groups indicates the users' interests and backgrounds. Also, the group statistic ismore
easy to obtain compared with other privacy concerning information, e.g., searching
history, the query log, etc. Therefore, we measure the affinity relationship between
user
u
m
and
u
n
using the cooccurrence of their joined groups:
n
(
u
m
,
u
n
)
W
m
,
n
=
(2.15)
n
(
u
m
)
+
n
(
u
n
)
where
n
(
u
m
)
is the number of groups user
u
m
joined and
n
(
u
m
,
u
n
)
is the number of
groups
u
m
and
u
n
co-joined.
Image affinity graph W
I
To measure the visual similarities between images, each
image is extracted a 428-dimensional feature vector
d
as the visual representation
[
20
,
47
], including 225-d blockwise color moment features, 128-d wavelet texture
features, and 75-d edge distribution histogram features. The image affinity graph
W
I
is defined based on the following Gaussian RBF kernel:
.
2
2
I
W
I
m
,
n
=
e
−||
d
m
−
d
n
||
/˃
(2.16)
˃
I
is set as the median value of the elements in
W
I
.
Tag-affinity graph W
T
where
To serve the ranking-based optimization scheme, we build
the tag-affinity graph based on the tag context and semantic relevance. The context
relevance of tag
t
m
and
t
n
is simply encoded by their weighted cooccurrence in the
image collection:
.
n
(
t
m
,
t
n
)
t
m
,
n
=
(2.17)
n
(
t
m
)
+
n
(
t
n
)
For tag semantic relevance, we follow Liu et al. [
20
] approach and estimate the
semantic relevance between tag
t
m
and
t
n
based on their WordNet distance:
2
·
IC
(
lcs
(
t
m
,
t
n
))
t
s
m
,
n
=
(2.18)
IC
(
t
m
)
+
IC
(
t
n
)
where
IC
is their least common
subsumer in the WordNet taxonomy. The tag-affinity graph is constructed as:
(
·
)
is the information content of tag, and
lcs
(
t
i
,
t
j
)
W
m
,
n
=
ʻ
c
t
m
,
n
+
ʻ
s
t
s
m
,
n
(2.19)
where
ʻ
s
are the weights of context relevance and semantic
relevance.
6
Note that we have no requirements on how to build the affinity graphs
and other intrarelation measurements can also be explored.
ʻ
c
+
ʻ
s
=
1,
ʻ
c
and
6
In the experiment, we choose
ʻ
=
0
.
9and
ʻ
=
0
.
1.
c
s
